| Paper: | MLSP-P7.10 (ICASSP 2003 Paper) | ||
| Session: | Pattern Recognition and Classification II | ||
| Time: | Friday, May 21, 15:30 - 17:30 | ||
| Presentation: | Poster (ICASSP 2003 Presentation) | ||
| Topic: | Machine Learning for Signal Processing: Bayesian Learning and Modeling | ||
| Title: | A RECURRENT MULTISCALE ARCHITECTURE FOR LONG-TERM MEMORY PREDICTION TASK | ||
| Authors: | Stefano Squartini; University of Ancona | ||
| Francesco Piazza; University of Ancona | |||
| Amir Hussain; University of Stirling | |||
| Abstract: | In the past few years, researchers have been extensively studyingthe application of recurrent neural networks (RNNs) to solvingtasks where detection of long term dependencies is required.This paper proposes an original architecture termed theRecurrent Multiscale Network, RMN, to deal with these kinds ofproblems. Its most relevant properties are concerned withmaintaining conventional RNNs’ capability of informationstoring whilst simultaneously attempting to reduce their typicaldrawback occurring when they are trained by gradient descentalgorithms, namely the vanishing gradient effect. This isachieved through RMN which preprocesses the original signalseparating information at different temporal scales through anadequate DSP tool, and handling each information level with anautonomous recurrent architecture; the final goal is achieved bya nonlinear reconstruction section. This network has shown amarkedly improved generalization performance overconventional RNNs, in its application to time series predictiontasks where long range dependencies are involved. | ||
| Back | |||
Home -||-
Organizing Committee -||-
Technical Committee -||-
Technical Program -||-
Plenaries
Paper Submission -||-
Special Sessions -||-
ITT -||-
Paper Review -||-
Exhibits -||-
Tutorials
Information -||-
Registration -||-
Travel Insurance -||-
Housing -||-
Workshops